
Nuclear scientist @Google building ML infra for Gemini, @GoogleDeepMind and @Waymo. Ph.D. @ImperialCollege. Steward of Gemini models. 🇬🇷🇨🇭🇬🇧🇷🇴🇶🇦
Zurich, Switzerland
Joined December 2011
- Tweets657
- Following357
- Followers364
- Likes1.5K
Pinned Tweet
How do we know if an AI agent is ready for the complex, adversarial realities of the real economy?
Traditional benchmarks test LLMs on static games (like Chess or Poker). But in the wild, agents face fluid, unseen environments where static leaderboards saturate and data contamination runs rampant.
Gemini 4 Carbon was the model that marked a step-change in how I approach research and development. Take it for a spin!
Dr. Alex Kenich retweeted
AI is a cheat code for ADHD brains.
You can jump between 10 projects, follow every rabbit hole, and somehow still get shit done.
I've never been more excited to build stuff.
Dr. Alex Kenich retweeted
Gemini 4 Argon is a surprise on transcription and it requires more analysis and explanation than normal.
Headline is that its’s clearly SOTA: modified WER of 1.73% and CER of 0.64% on low effort. But that’s not the whole story because essentially the benchmark is not saturated.
First of all, Argon works a bit differently. It follows instructions better (it successfully differentiates and ignores archival markings, page numbers etc) and it also does better with unusually formatted pages and text, like the attached image. Most importantly its errors are also far less consequential… and I’m not sure we should call them errors.
When the initial results came back, they felt much better than the raw numbers suggested. At first it was scoring about the same as 3.1 pro but the transcriptions simply looked better. So we did a deep dive and when we examined all 174 word errors, we found that all but 40 were essentially toss-ups. These were all cases where the original was misspelled but where it’s also unclear what the correct, ground truth misspelling should be. Basically squinting and arguing about whether that clearly incorrect letter is an erroneous e, o, or a. The point is these do t really matter because the original is unclear.
This hadn’t really been an issue on previous tests because most of the error were still meaningful or at least represented things like a misspelled word in the original that had been corrected. This was different.
When you put it all together, we basically have a saturated benchmark, at least on 18th and 19th century English text. Gemini 4 Argon’s effective word error rate is something like 0.4% and it’s effective CER is about 0.2%. I am being imprecise because people will disagree about how we should handle these types of residual errors.
What is clear, is that on our benchmark anyway, Argon is much better than human experts. It’s also slightly cheaper to use than 3.1 pro and even a bit faster, in our testing.
This is a significant moment because basic textual accuracy is the precursor to everything else historians want to do with documents (and any other knowledge work which involves historical documents or handwritten papers). Once you have accurate text, you can start to actually envision scaling out massive research projects.
And it looks like we’re there.
Sorry this was a bit late (people have been asking), it’s been a very busy week and I wanted to get this right.
We post-trained Argon exclusively on my rolled back CLs to top this benchmark
GEMINI 4 ARGON JUST WON THE ONE BENCHMARK NOBODY EXPECTED 💀
Forget coding.
Forget reasoning.
Forget math.
GEMINI JUST TOOK #1 FOR HUMOR. 😂
Artificial Analysis’ new AA-PersonalityBench tested 36 AI models on how well they use humor, wit and playful language.
And Gemini 4 Argon came out on top with 1573 Elo 👀
Ahead of Claude, GPT, Grok, DeepSeek and the rest.
And honestly…
If you’ve used Gemini, this result probably makes perfect sense 💀
Sometimes you ask a completely serious question and Gemini decides:
“Let me make this funny first.” 😭
But this benchmark actually raises an interesting question:
We spend so much time asking which model is the smartest…
What if the best AI is simply the one you actually enjoy talking to? 👀
So be honest:
Is Gemini the funniest AI right now? 😂
1
55
7,079
Basic supply and demand: AI is flooding the market with expertise. Likeability is the new scarce good. The smart jerks are about to learn about comparative advantage the hard way.
1
78
Dr. Alex Kenich retweeted
We are bringing agents to the wet-lab bench at Google Research to navigate complex biological landscapes. Learn how we are grounding agents in physical lab robotics to run fast closed-loop molecular design cycles, ultimately discovering single-digit nanomolar protein binders for cancer targets →goo.gle/4B5mMMB
Just looked through some decompiled code because I'm mad and noticed that agents tend to leave a LOT of performance gains on the table. For now, agents tend to make suboptimal choices when reverse engineering binaries. As obvious as it sounds... Try asking them explicitly to optimize by:
1) using guard clauses
2) batching syscalls and heavy I/O operations
3) reusing memory instead of reallocating
4) warming up / precomputing immutable tables and configs just once at initialization
5) using optimized data structures (e.g. don't do lookups in massive arrays)
6) running performance benchmarks on changes
7) identifying independent loops and parallelizing the heck out of them
8) identifying opportunities to cache, and doing it generously
pro tip: a lot of older software was built to run on hardware where compute and memory was far more constrained than today, so performance trade-offs had to be made. Make sure your agent knows your machine's hardware specs so it can take full advantage.
2
122
Preach. A lot of these results are looking like we're hitting the floating point precision artifacts of the universe
I've been making my own audiobooks of my favourite classics with this; full casts each with their own expressive voices. No more lame narration and guys terribly voicing female characters
introducing Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, our most expressive audio generation models yet
these models enable creators, developers, and enterprises to create richer, more expressive audio experiences
try them via the Gemini API and in AI Studio: aistudio.google.com/generate…
2
134
Dr. Alex Kenich retweeted
trained a fly connectome to handle kubernetes deployments
imagine escaping your mortal body only to find out you're now primary oncall
For the first time, scientists have mapped the complete brain and central nervous system of an adult male fruit fly — a key model organism in science. 🪰
Working alongside HHMI Janelia Research Campus and the scientific community, @GoogleResearch scientists and researchers used AI to combine millions of 2D images into 3D neural shapes, reconstructing a record-breaking 166,000+ neurons. This foundational map of the adult male fruit fly brain can help accelerate our understanding of the brain, and is a major milestone in neuroscience.
Okay, this is the best way to explain why AI Slop is not what you think.
By using AI Denzel Washington to explain it all. This is actually really good. Well written, directed and generated.
It's about you not needing the studios anymore.

























